Notes

An AI Internal Knowledge Base, Answers From Your Own Documents

What is an AI internal knowledge base?

An AI internal knowledge base is an assistant that answers your staff's questions in plain English using your own documents: standard operating procedures (SOPs), policies, the handbook, price lists, product sheets, the how-to that lives in a shared drive, SharePoint, Google Drive, Confluence or Notion. It finds the relevant passage, writes the answer, and shows where the answer came from so the person can check it.

The point is not a smarter search box. The point is that the question "how do we handle a return without a receipt" gets the answer your business actually uses, in ten seconds, without interrupting the one person who knows.

Why not just use a general chatbot?

Because a general chatbot answers from the internet, and your staff need the answer from your business.

Ask a public tool how to process a faulty goods claim and you will get a reasonable sounding generic answer. It will not know your supplier, your approval limit, or that you changed the process in March. A grounded assistant is restricted to the documents you give it, and when those documents do not cover the question, the right behaviour is to say so rather than improvise. That difference is most of what we mean when we talk about AI hallucinations in business tools: an answer without a source is a guess.

How does it give grounded answers with sources?

It retrieves the passages that match the question from your documents first, then writes the answer only from those passages, with a link back to each one.

Under the hood, your documents are broken into sections and indexed by meaning, which is what people mean by semantic search, so a question about "leave for a funeral" finds the policy headed "compassionate leave". The assistant reads the few sections that match and answers from them. Every answer carries its sources as citations, and staff should be taught to click them for anything that matters. A source they can open is what turns a confident sentence into a trustworthy one.

Good builds also log the questions that found nothing. That list is gold: it is a map of the knowledge gaps in your documentation, written by the people who hit them.

Who should be able to see what?

Employees should only get answers from documents they would already be allowed to open, and this has to be decided before anything is connected.

The usual failure is quiet. Someone connects the whole shared drive, and suddenly a casual on the floor can ask about the payroll spreadsheet or a performance file because it happened to sit in a folder. Permissions and access controls need to follow the person asking, not the assistant. If your current folders are a mess of "everyone can see everything", fix that first, because the assistant will faithfully expose whatever mess it is given.

Personal and other sensitive information needs the same care. The OAIC guidance on privacy and commercially available AI products is a sensible read before you choose a tool, and the ACSC has practical advice on securing the systems the documents sit in. How those obligations apply to your business specifically is a question for your adviser, so treat this as a starting point.

What documents do you need first?

The documents your staff already ask about, in their current, correct version, and nothing else to begin with.

Start by listing the twenty questions that interrupt someone every week. Then find the document that answers each one. You will discover three kinds of result: the answer exists and is current, the answer exists in four versions that disagree, or the answer is not written down anywhere. Only the first kind is ready. The work of fixing the other two is the same work described in preparing your data for AI, and it is worth doing whether or not you build anything.

Resist the urge to load every file you own. Old drafts, superseded policies and personal notes make answers worse, not better.

How do you keep the documents current?

Give every document an owner and a review date, and make the assistant's answers show how old their source is.

An assistant that answers confidently from an out of date policy is worse than no assistant, because it sounds certain. The fix is ordinary housekeeping: a named owner per document, a review cycle that matches how often the thing changes, and a simple feedback button for staff to flag an answer as wrong. If your documents live in a system with version control, the assistant should always read the current version and ignore the rest. When a process changes, the document changes the same day, or the assistant keeps teaching the old way.

The questions that found nothing, and the answers staff flagged, should land in front of the document owners on a regular schedule. That is how the knowledge base gets better instead of slowly drifting.

Where does it earn its keep?

It earns its keep wherever one experienced person is being interrupted all day by questions that have a written answer.

New starters are the obvious case, and we cover that in AI for employee onboarding. Multi-site businesses are another, where the same process question is asked at every location and the answers sit in data silos, one folder per site. Customer-facing teams who need the right product detail while someone waits on the phone are a third. In each case the measure is simple: how often the senior person is interrupted, and how often the answer given was right.

A practical first step

Write down the twenty questions. Find the documents that answer them. Fix the ones that disagree. That alone will help, and it makes a grounded assistant a small, sensible build rather than a gamble.

If you would like it built on your own documents with permissions done properly, see agents on your data, or get in touch and tell us which questions your team asks most.

Keep reading